| Message ID | 4D9E37A4.8040902@seznam.cz (mailing list archive) |
|---|---|
| State | Rejected, archived |
| Headers | show |
diff --git a/memops.c b/memops.c index 45bd340..328edec 100644 --- a/memops.c +++ b/memops.c @@ -126,7 +126,7 @@ static void simplify_loads(struct basic_block *bb) if (!dominators) { if (local) { assert(pseudo->type != PSEUDO_ARG); - convert_load_instruction(insn, value_pseudo(0)); + convert_load_instruction(insn, insn->src); } goto next_load; }
Provided that patch by Nicolas Kaiser (commit 5ecad11 in Chris Li's repo) is correct (I did some work of comparing resulting code with and without this change and some unnecessary loads seem to be additionally eliminated with that change applied [desirable]), another problem has shown up that the always-true conditional used to mask. This is an example case*) that is not proceeded correctly: static int test(void) { int x; int *px = &x; return *px; } Original result: <entry-point> # symaddr.32 %r1 <- x # snop.32 VOID -> 0[px] # lnop.32 %r2 <- 0[px] # lnop.32 %r3 <- 0[x] # phisrc.32 %phi1(return) <- VOID # phi.32 %r4 <- VOID ret.32 $0 Result with this patch on: <entry-point> # symaddr.32 %r1 <- x # snop.32 VOID -> 0[px] # lnop.32 %r2 <- 0[px] # lnop.32 %r3 <- 0[x] # phisrc.32 %phi1(return) <- VOID # phi.32 %r4 <- VOID ret.32 x Originally, I suspected that similar change should be made also to flow.c (in `find_dominating_stores'), but my experiments made me believe that this case is right as is -- it starts to do less optimizations after such edit (encountered, e.g., on tokenize.c with bitfields handling in `stream_pos'). My effort to check that nothing goes wrong with this patch (comparing against 0a782d3 commit in Chris Li's repo) was following: 1. checking both the warnings and linearized codes**) for C sources in linux-2.6.35.12/kernel dir - warnings: the same (the same also with Nikola Kaiser's change removed) - linearized codes: the same (compared with the state without Kaiser's patch, there was a few differencies: compat.c, exit.c, mutex-debug.c, sys.c, time.c, tsacct.c; these seem to be unnecessary loads as mentioned at the beginning) 2. checking linearized codes of sparse itself -- contrary to 1., this was a bit difficult because sparse generates different in-code addresses (at least it did for me) so I had to find out the way how to "similarify" them before comparison***) - the only change I encountered is a "strange" instructions swap which I currently have no explanation for (maybe it is due to some sorting involved; the only thing I know is that the order of instructions is the same as it is without Nikola Kaiser's change): --- orig/sort.c.log 2011-04-07 17:12:47.000000000 +0200 +++ with_patch/sort.c.log 2011-04-07 17:15:36.000000000 +0200 @@ -477,8 +477,8 @@ sort.c:152 # phisrc.32 %phi50(b2) <- %r155(b2) # phisrc.32 %phi68(i1) <- %r90(i1) # phisrc.32 %phi72(i2) <- %r99(i2) - add.32 %r114(nbuf) <- %r201, $1 muls.32 %r115 <- %r201, $4 + add.32 %r114(nbuf) <- %r201, $1 add.32 %r116 <- %r115, buffer br %r110, .L0xb76fe764, .L0xb76fec8c *) the other thing is that no warning about a use of uninitialized value is emitted (neither by library nor by sparse binary); the same applies to case when there is plain "int x; return x;" **) using commands like this: $ cd linux-2.6.35.12 && mkdir sparse_out $ for i in $(find kernel -name *.c); do make C=1 CHECK=<path to sparse> \ CF="-ventry -vv > sparse_out/$i.log" $(echo $i | sed "s/\.c/\.o/"); done ***) I wrote a small, stupid and slow Python script for this purpose, you can find it at https://github.com/jpokorny/thesis/blob/master/cl_sparse/addr_reg_adj (if there is an interest, I am not against including it into sparse repo, maybe anyone else would use it for smooth linearized code comparison, or maybe I am missing something about how to test unwanted changes in linearized code?) Signed-off-by: Jan Pokorny <pokorny_jan@seznam.cz> --- memops.c | 2 +- 1 files changed, 1 insertions(+), 1 deletions(-)